A diagonal covariance model gives mean high-gamma AUC 0.5335 for random labels when the reference excludes positive contacts in 70 HUP SEEG samples

We derived an exact expected area under the ROC curve (AUC) for an average reference that excludes positive contacts. The diagonal covariance model permits unequal contact powers. It requires positive powers, at least three negative contacts, and two label classes. We applied it to 70 one-minute interictal SEEG samples from 35 patients in OpenNeuro ds004100, version 1.1.3. With uniform random labels at each run's clinical positive count, the exact mean model AUC was 0.533486. A fixed reference has expected AUC 0.5 under this rule. The proof also shows that exclusion cannot lower AUC under the model assumptions. With measured covariance and 4,096 draws per patient, mean AUC rose from 0.500140 to 0.508317. The paired increase was 0.008178, with 95% patient bootstrap interval [0.006114, 0.010702]. A frozen clinical reference gave mean AUC 0.500133. The measured effect confirms concurrent AutoSci work in the same 70 samples. This is not an independent cohort replication. The exact unequal-power formula and per-run count preservation are added contributions. The package also has frozen-reference and within-shaft controls. All 70 fresh public signal prefixes matched their saved hashes and covariance values. An independent EDF reader checked one run. Independent pair calculations checked 6,160 AUC values. A full repeat gave 75 byte-identical analysis files. The package preserves the proof, code, compact data, results, provenance, and pinned raw-data retrieval recipes. This study does not establish clinical prediction accuracy or a neural mechanism.

stereoelectroencephalography · average reference · high-gamma power · area under the ROC curve · random labels · diagonal covariance · reference exclusion · epilepsy